In this work, we select three public video datasets commonly used for violence recognition and crowd behavior understanding: Real-Life Violence Situations (RLVS), RWF–2000, and HockeyFight. These datasets include aggressive behaviors such as interactions with crowds, crowd provocation, and non-violent actions in difficult, uncontrolled real-world settings.

For this study, we chose three widely-used public video datasets that focus on violence detection and crowd behavior analysis: Real-Life Violence Situations (RLVS), RWF-2000, and HockeyFight. These datasets contain various forms of aggressive conduct, including crowd interactions, instances of crowd provocation, and non-violent activities, all captured in challenging and uncontrolled real-world environments.